Evaluation of HEVC decoders (SW, Hybrid and HW)

[NikosD]

Die you try to compare the Videoprocessing (Luma and chroma scaling, Deinterlacing)?
Maybe in YV12 you can not use the adaptiv GPU Deinterlacer (bob/weave Fallback) or the "good" GPU scaler (bilinear fallback).
I think Intel should reply to all these.
What exactly are the differences of NV12 and YV12 regarding scaling algorithms or any other reason of slow performance.

Because for Nvidia and AMD, I haven't seen such differences.

Also disable all "enhancements". For example Skin Tone Enhancement and Contrast Enhancement.

I always do that.

Actually, it's the first thing I do for all my GPUs (Nvidia, AMD, Intel)

In Intel forums, I even suggested them to put everything off by default, but Intel thinks those "enhancements" are useful and tested by them.

[Kevin Gilmore]

Is it possible to 'pre-render' high-bitrate HEVC content
through Broadwell (or 10bit HEVC via Skylake:) Hybrid decoders such that I/one can then watch forthcoming/the future 4kBlurays (more high bit-rate HEVC/VP9 content) through my hybrid decoder solution in real time?

That is, Haswell/Broadwell/even SKYLAKE is only 'good' at best for real-time decoding at say ~18Mbps, given their respective formats (and shader faculties).

Put another way, or again......, I'd like to decode ~100Mbps or even say up to 600Mbps .265 files in advance, so that I can get more use than compressed youtube 4k from my Hybrid-solution decoders?
For instance, we'd desire to use a 'quasi-new Broadwell', and "new" Skylake chips to run my 4k TV or monitor at real-time rates, can we[B] 'pre-render/process' (I'd hope not transcode where Sky's capable!) Hbrfiles a la solution [for 'real-time' playback]?


It appears that Broadwell non-GT3e can't hybrid decode 10bit according to Intels' IGP driver.
People here have posted Broadwell Gt2 can indeed decode HEVC8/10 and VP9 though and I've seen references before-made that 5th gen 2+2 has been decoding VP9 (through YT HTML5 player) for this past year...

Some type of Processing or rendering anteceding playback would be a terrific H-br .265 'solution' for those without whole ff decoders (or say desktop gm206) until our next -particularly mobile- upgrade, (perhaps in Broadwell or Haswell/Skylake user case sometime in 2017/2018) etc.

Seeing as 6th/7th gen probably bring nothing else substantial to the table than Fixedfunction decoders - if Hybrid decoders could pre-process.... HBM, better OLED/DCPI3 =>4klaptop displays, and (noiseless) lower TDPs' are just around the corner.. Mentioned because they are very much relative to MadVR/LAV experiences and rendering..
6th/7th gen with accompanying GPUs' are going to be initially expensive circa release. I'm just trying to, "have the cake and eat it" now/today (as opposed to early thru late next year at best-case scenario in the mobile Form-factor) while prosaically saving a buck, with the hybrid decoders we already own.

Is preprocessing UHD-esque or High bitrate .265 possible for proper hybrid decoder playback? Appreciate and have empathy for those who read this WoTxt.

I've a Broadwell 5200U/940M if I can be helpful/run any tests..

[huhn]

not going to happen

first of all it would take hours to decode and the size of that decode file needs to be on a SSD.
a HDD is way to slow and to make things even worse the SSD needs to be huge too...

so forget it. HEVC hybrid decoders are kind of useless.

[vivan]

Yeah, 10bit 2160p@60 fps is 3840*2160*15*60 = 7 Gb/s. Which is around 50 GB/min.
Your only way is reencoding.

[leonccyiu]

it's not like you have to play the VP9 version.

That's true, but 4k60 videos are only encoded in vp9 and only show 1080p60 max in internet explorer. At the moment 4k60 videos are few, but as time passes they may become more common.

I also do notice that vp9 is in general less prone to artifacts and I think it is a superior compression algorithm. Whereas VP9 is used to reduce bit-rates for 1080p and below, it seems for 4k videos, many of them have higher bitrates, sometimes significantly so than h264.

I can't wait for full hardware acceleration to be working. Chrome seems to be worst at implementing it, even though on chrome://gpu everything is checked as working, when I use BBC iplayer or anything streaming h264 on html5 I can see cpu usage rise corresponding to video quality meaning hardware accelerated decoding is not working.

[NikosD]

I would call that a miracle, but as you know miracles have stopped happening the last few thousand years.

Latest Intel driver v4300 for Haswell Win 10 and better have "fixed" the performance drop of NV12 (!)

Now, using DXVA Checker and MPC-HC, I see no performance difference between NV12 and YV12 (!!)

BUT looking carefully at the picture of decoded clip, I see the same artifacts using NV12 like the ones using YV12.

Before v4300, the picture quality of NV12 was better.
Now it matches YV12, like performance.

This must be wrong, but it seems that Intel reacted insanely fast and changed the NV12 to YV12, but when ?

I reported that issue on October 25 and the drivers were ready by October 1-16 (maybe not closed completely, however)

Actions needed:
@Nintendo Maniac 64
@Wanezhiling
@leonccyiu

Can you test my findings ?

[wanezhiling]

There is no v4300 driver for win8.1 and so I can't test...

[foxyshadis]

BUT looking carefully at the picture of decoded clip, I see the same artifacts using NV12 like the ones using YV12.

Before v4300, the picture quality of NV12 was better.

What do you mean? Decoding bugs? Do you have any screenshots?

[NikosD]

I copy my comment from the other thread.

I was talking about the 4K VP9 clip I uploaded and some artifacts that were visible with YV12 and invisible with NV12.

Read here:

It is true that when the camera zooms in from 28 sec to 39 sec, the picture in windows mode looks like a little weird around the grass, like transparent water floating around.

In full screen (1920x1080), everything looks good.

[P.J]

Isn't the performance difference between those colorspaces because of their scaler method?

[nussman]

Latest Intel driver v4300 for Haswell Win 10 and better have "fixed" the performance drop of NV12 (!)

Now, using DXVA Checker and MPC-HC, I see no performance difference between NV12 and YV12 (!!)

BUT looking carefully at the picture of decoded clip, I see the same artifacts using NV12 like the ones using YV12.

Before v4300, the picture quality of NV12 was better.

I would call this a bug and not a fix .. :eek:
Disabling "good" processing for playback FPS benchmarks? Come on ...
I hope you dont "push" Intel to do something stupid ...

[NikosD]

It's not exactly playback benchmark.
It's plain playback, just measured in numbers.

The difference is huge, it's like you can play it in realtime or not (dropped frames or slow performance)

I've never told Intel to increase the performance by dropping the quality.

And between you and me, who am I to tell Intel anything ?

But we need more people to confirm or deny my findings.

[leonccyiu]

I would call that a miracle, but as you know miracles have stopped happening the last few thousand years.

Latest Intel driver v4300 for Haswell Win 10 and better have "fixed" the performance drop of NV12 (!)

Now, using DXVA Checker and MPC-HC, I see no performance difference between NV12 and YV12 (!!)

BUT looking carefully at the picture of decoded clip, I see the same artifacts using NV12 like the ones using YV12.

Before v4300, the picture quality of NV12 was better.
Now it matches YV12, like performance.

This must be wrong, but it seems that Intel reacted insanely fast and changed the NV12 to YV12, but when ?

I reported that issue on October 25 and the drivers were ready by October 1-16 (maybe not closed completely, however)

Actions needed:
@Nintendo Maniac 64
@Wanezhiling
@leonccyiu

Can you test my findings ?

Playing the LG Garden Demo and Panasonic Angkor Wat demo from demouhd3d, from the MADVR osd, it says that they are NV12 and they are playing back absolutely fine with no image artifacts whatsoever. The video looks a little washed out compared to what I remember in the showroom, but that's because of the poor contrast of the ips monitor I have.

The Youtube 4k vp9 videos are all YV12, I do not notice any major quality differences on my 1080p monitor.

[NikosD]

The test shall be done by disabling NV12 from LAV Video properties and then YV12 is used.

You could then compare, after you disable NV12, the performance and quality of YV12 compared to NV12 before and after v4300 drivers.

You shall use a 4K scaled to a smaller window video, in order to use the scaling algorithms.

That's it!

[nevcairiel]

And don't use madVR, it behaves entirely differently and is probably not affected by the same quirk. Only EVR.

[huhn]

just to make this more clear.

is only downscaling affected or is up scaling affected too?

[CruNcher]

I5-2400 Sandy Bridge 4 Cores Multithreaded Fluid Playback Results Lav Video (OpenHevc) CPU

8 Bit:

4K-Changhong-Furano-HEVC-HD.Club.mp4
4K-TCL-super_car-Part-1-@HD.Club.mp4
Demo Ultra-HD - TCL - 4K HEVC.mp4
LG_4K_View-the-Feeling.mp4
Hispasat_UHD_NAB-2014_HEVC.mkv
Beauty_3840x2160_120fps_420_8bit_HEVC_MP4.mp4

10 Bit:

Exodus_UHD_HDR_Exodus_draft.mp4
Samsung_Fashion_Show_Milan_UHD.ts (Very near of the Edge)
Samsung_SUHD_Tech_Story_Europe.ts
Samsung_UHD_Dubai.ts
UHD_ENT_Transformer_Quad.ts
UHD_PD_Tech_Story_Euro.ts
UHD_PQ_Dive_to_New_World.ts
UHD_PQ_Iceland_2.ts
UHD_PQ_The_City_Limits.ts


Close 10 bit fail:

Shut.Up.And.Play.Berlin.2014.ts (Very close 3 High Motion Scene glitching out, still counted as a fail)

Interesting Results:

Eutelsat.Demo.HEVC.ts <- They made a nice Artistic Chaotic Motion Test Sequence :)

[wanezhiling]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 3.10.0 (2015/11/07)
Added "MJPEG_VLD_PureVideo" as an alternative name of Decoder Device
Minor changes and fixes

http://i.imgur.com/ONJTooI.png

[NikosD]

http://bluesky23.yukishigure.com/en/DXVAChecker.html


By running the "Select Format" option and then selecting "Motion JPEG", is there any HW MJPEG decoder installed for Nvidia ?

Intel's driver installs "Intel Hardware M-JPEG decoder MFT" which is working for Haswell and it is a very fast HW decoder.

You could also check directly by using "Check Decoders" -> "Select Media File" and choosing a MJPEG video file.

If Nvidia installs a working HW MJPEG decoder, I could upload a 1080p and a 4K sample to test it.

[huhn]

http://abload.de/img/mjpegfwr9k.png

and it works.

edit: a short test shows the video engine isn't used for decoding. decompression JPEG or something like that is most likely a normal task for a GPU.

[NikosD]

and it works

OK.

Using Win 10 x64 - Core i7 4790 - iGPU HD 4600@1.5GHz - v4300 and Intel MJPEG HW decoder x64 with DXVA Checker v3.10.0 (in pure decoding mode) I got these results:
(CPU usage <10%, GPU usage 0%)

The first clip is 1080p50fps@105Mbps
Here:
https://www.sendspace.com/file/hfhw5r

For the first clip, I needed to put "Repeat Count" at least 5 times, because decoding was too fast.

Clip 1:
Intel MJPEG HW decoder x64 1584/1585/1584


The second clip is 2160p50fps@292Mbps
Here:
http://www.filedropper.com/crowdrun4knoqjpeg

Repeat count 5.

Clip 2:
Intel MJPEG HW decoder x64 401/402/401

[huhn]

looks like both ffdshoiw DXVA1 and NVDIA MJPEG doesn't support MKV. AVI works fine.

[P.J]

Please don't upload on filedropper =/

[huhn]

yeah that took some time to DL..

[huhn]

i remuxed the files and here are the results...
3770k 3.5 GHZ
960 GTX 1430 mhz
tested driver 358.87

decode mode.

UHD doesn't work.

1080p:

12% CPU usage
DXVA checker doesn't see the GPU usage
so looks like it is limited by the CPU.
200 - 210 -215
GPU-z says core 135 mhz 202.5 memory clock

about 55% GPU usage at that speed and the memory controller is at ~35%
so something is happening at the GPU not sure what...

this is either new or they don't care.

[NikosD]

Please don't upload on filedropper =/

It is the easiest way for me to upload > 300 MByte files.

Do you know any other file server working with a simple drag and drop for uploading > 300 MByte files ?



UHD doesn't work.



It should work according to DXVA Checker.

So, probably the 292 Mbps bitrate is forbidden for Nvidia's hybrid MJPEG decoder.


1080p:

12% CPU usage
DXVA checker doesn't see the GPU usage
so looks like it is limited by the CPU.
200 - 210 -215
GPU-z says core 135 mhz 202.5 memory clock

about 55% GPU usage at that speed and the memory controller is at ~35%
so something is happening at the GPU not sure what...

this is either new or they don't care.

The 1080p clip is still a 105 Mbps clip, so CPU is probably the bottleneck due to that bitrate.

Anyway, MJPEG is a QuickSync HW accelerated format for Intel, so is unbeatable by hybrid or CPU decoders.

[P.J]

i remuxed the files and here are the results...
3770k 3.5 GHZ
960 GTX 1430 mhz
tested driver 358.87

decode mode.

UHD doesn't work.

1080p:

12% CPU usage
DXVA checker doesn't see the GPU usage
so looks like it is limited by the CPU.
200 - 210 -215
GPU-z says core 135 mhz 202.5 memory clock

about 55% GPU usage at that speed and the memory controller is at ~35%
so something is happening at the GPU not sure what...

this is either new or they don't care.

Same here. LAV uses only one thread too.
I have never seen any MJPEG video...

[nevcairiel]

I have never seen any MJPEG video...

Some older home-video cameras recorded in that, but its otherwise not very common, indeed.

[P.J]

Some older home-video cameras recorded in that, but its otherwise not very common, indeed.

But CPU can handle 1080p MJPEG and even 4k well.

[NikosD]

Its a shame they didn't build an official DXVA2 specification for VP8/9 decoding, or released their proprietary way to do it (as it clearly goes through some DXVA2 device), oh well. Only being able to use MediaSDK and not implement it using existing DXVA2 decoders is sad.

After reading the above post I had sent an email to Intel - on August 19.

I got an official reply today that:

"VP9 dxva is already enabled at driver level.
VP8 dxva will be enabled post-Win10/2.
We(MSDK) don’t plan to provide DXVA MFT.
MSFT will provide them in OS at the next revision time frame."

Also, they told me that Haswell is not going to support VP8/VP9 in any way.

They are not aware of any ISV’s that are currently using those MSFT GUIDs, but I told them that you (the developer of LAV filters) are working on this.

Am I right ? What is your progress ?

[nevcairiel]

VP9 DXVA works on Braswell/Skylake with my code now. NVIDIA exposes the GUID as well, but thats still broken in recent drivers.
Didn't do anything for VP8 yet due to lack of testability. Once driver support arrives.

[NikosD]

VP9 DXVA works on Braswell/Skylake with my code now.


Is it hybrid for those two architectures ?

[nevcairiel]

From what I understand yes, Gen9 GPUs only have hybrid VP9 decode.

[NikosD]

OK, thanks. It makes sense.

But I didn't see VP9 DXVA in the nightlies.

Is there something holding you back of releasing the VP9 DXVA decoder ?

[Alecex]

So am I getting this right? The GTX970 has almost no positive effect on h.265 Decoding when compared to what the gtx960 is capable of?

Did anyone test the AMD Fiji Graphicscards or the Nano?

If the ATI r9 380 or the upcoming R9 380X (Antigua XT) supports h.265 4k 60p playback as well as the GTX960 does, then i would prefer the ATI beacuse its the better gameing card as well for the price.

The GTX 950 is supposed to be as good as the 960 for video but 60 usd cheaper so did someone confirm its h.265 capabilities?

[nevcairiel]

The GTX960 is superior to all the other cards you listed for HEVC/H.265 decoding, because all the other models either only implement 8-bit hybrid decoding (GTX970, AMD before Fiji), or lack 10-bit decoding (AMD Fiji).
On top of that, the GTX960 is the only of the lot that fully supports HDMI 2.0 and HDCP 2.2. GTX970 only has HDMI 2.0, but not HDCP 2.2, and AMD is limited to HDMI 1.4

[Alecex]

Thanks a lot!

I hope they release a GTX 960 TI at some point with 4gb and 256 bit interface...

its painfull to know how much more this Graphicscard could do with a proper interface.

Maybe one schould buy the gtx950 untill they release a 960 TI... instead of going directly for the 960. Or is there any limitations on the 950 video wise?

[nevcairiel]

There are GTX960 models with 4GB, which should be fine. The available memory bandwidth should be enough for any video usage.

There were rumors of a 960 Ti before, but that was in the beginning of this year, and since it hasn't shown up by now, its unlikely that it still will.
Pascal will hopefully (and quite likely) have the full decoder chip in its entire lineup, so that offers some new options then.

[Alecex]

Yes but with a 950 u save money if its as good as the 960 for Video, has anyone tested it? if not i could get a 950 first.

When is pascal expected to launch? I dont want to wate for Pascal my GTx460 is so loud and nerve-wrecking... plus it only does 1080p h.264 palyback/decoding. I wil get a pascal but untill then i need to change my gtx 460.

[CruNcher]

So am I getting this right? The GTX970 has almost no positive effect on h.265 Decoding when compared to what the gtx960 is capable of?

Did anyone test the AMD Fiji Graphicscards or the Nano?

If the ATI r9 380 or the upcoming R9 380X (Antigua XT) supports h.265 4k 60p playback as well as the GTX960 does, then i would prefer the ATI beacuse its the better gameing card as well for the price.

The GTX 950 is supposed to be as good as the 960 for video but 60 usd cheaper so did someone confirm its h.265 capabilities?

There is a effect that can be measured in a small energy saving with Nvidias Shader Decoder in GM204 but its relatively small compared to OpenHevc on the 4 Core I5-2400 and it can playback more complex streams and of course 10 bit that Nvidias Shader Decoder fails on not sure how this would look like if the Shader Decoder would operate on DX 12 (Win10) but on Windows 7 it's not really worth the hassle with the GTX 970 currently.

For low complex Streaming i would still overall prefer Nvidias Shader Decoder, because it is overall still more efficient compared to my I5-2400 (95W TDP) in Power Saving with such lower complex Online Bitstreams.

But the more closer you get to Broadcast Complexity the more it hits Nvidias Shader Decoder Efficiency vs the CPU alone and the CPU takes the lead here.

So the more efficient the CPU the weaker this will become and puff it's gone maybe already with the same region 4 Core Haswell CPU with Lower TDP on OpenHevc, this could be already the case.

I will still keep an eye if Nvidia is maybe able to improve their Shader Decoder furthe (if they should try it but why should they if they have VP8 ready ;) ) from what i measure i doubt this is possible to the degree it would need to conquer the 4 Core Sandy Bridge CPU OpenHevec Performance here anyways on the more Complex Broadcast Bitstreams ;)

So either GTX 960 or waiting for Pascal where the VP8 becomes the Standard Decoder for all Cards ;)

In either Way as you still want to play Games with good options on like 1080p 60 FPS neither the GTX 950/960 are sufficient today you need at least the GTX 970, especially for more complex GameWork (Compute) stuff like Hairworks or VXGI where the FPS can easily drop down to 1080p 30 FPS.

For normal WEB Online Bitstream Complexity your are fine even with the Nvidia Shader Decoder for stuff like for example 4K 10 bit Netflix you would need to fall back to the CPU or internal GPU if you have one that supports H.265 10 bit like Intels Gen9 Quicksync or AMD Carrizos UVD.

[NikosD]

So am I getting this right? The GTX970 has almost no positive effect on h.265 Decoding when compared to what the gtx960 is capable of?

Did anyone test the AMD Fiji Graphicscards or the Nano?

If the ATI r9 380 or the upcoming R9 380X (Antigua XT) supports h.265 4k 60p playback as well as the GTX960 does, then i would prefer the ATI beacuse its the better gameing card as well for the price.

The GTX 950 is supposed to be as good as the 960 for video but 60 usd cheaper so did someone confirm its h.265 capabilities?

I think nowadays you can't have it all.

Nvidia GTX 960 supports fully the H.265 (HEVC) codec, but lacks VP9 as all cards right now and it's not as powerful as R9 380 for gaming.

If you think about the future, meaning DX12 games, you should avoid buying a Nvidia card.
Maybe Pascal could be better, but right now Maxwell has the well known problems.

AMD on the other hand is behind Nvidia for HTPC since it doesn't support 10bit HEVC and HDMI 2.0, yet, although its cards are a lot better suited for DX12 gaming.

Intel has the same problems with AMD, lack of 10bit HEVC and you will definitely need a discrete card for gaming, although it has the faster decoder/encoder for H.264 and 8bit HEVC.

On the other hand, the content in HEVC is so little.

I think right now VP9 is more important than HEVC, since YouTube using Chrome uses that codec by default and sometimes explicitly (4K60fps clips)

If I were you and wanted desperately to buy something right now, I would buy an AMD card if DX12 gaming was my first priority or I would wait for next year's cards from both AMD and Nvidia that will probably have everything accelerated in HW (AVC, HEVC (8bit/10bit), VP9) and will certainly be faster in DX12 gaming than this generation.

[P.J]

I think right now VP9 is more important than HEVC

Satellites are broadcasting 4K in H.265 10bit format. 4K Blu-rays will hit the market soon.
Not all people have fast internet to watch 4K VP9, and it's useless for streaming 4K in low bitrate, could be much worse with fake upscaled videos too.
People can download the source video which has better quality than YouTube 4K.

And there's no proper DX12 game right now
http://www.anandtech.com/show/9740/directx-12-geforce-plus-radeon-mgpu-preview/4

[NikosD]

Satellites are broadcasting 4K in H.265 10bit format.


That's the case for very few people watching 4K H.265 DVB-S.


4K Blu-rays will hit the market soon.


There will be no 4K Blu-rays for 2015.
4K Blu-Ray will start sometime in 2016, but massively we don't really know.
Probably 2017 or even later


Not all people have fast internet to watch 4K VP9, and it's useless for streaming 4K in low bitrate, could be much worse with fake upscaled videos too.
People can download the source video which has better quality than YouTube 4K.


ALL YouTube videos and not only 4K are in VP9 format right now, when Chrome is used.
It's not that VP9 is used for 4K only.


And there's no proper DX12 game right now
http://www.anandtech.com/show/9740/directx-12-geforce-plus-radeon-mgpu-preview/4


As I said, DX12 gaming is a future (2016) thing too.

[P.J]

Recent CPUs can decode 1080p well but 4K.

[rodrigo081089]

Hi, i bought the gtx 960, i can reproduce 4k 8bit h264 videos ok, but when i tried 4k 10bit hevc i have a lot of dropped frames.

I use the MPC, could it the MPC reproduce HEVC ok?

[huhn]

make sure you are using DXVA not CUVID

[rodrigo081089]

make sure you are using DXVA not CUVID

If i use DXVA even the 1080p videos are unwatchable, i had a drop frame every second

[huhn]

than there is something wrong with your system. i can only guess but this is the wrong thread for this.

crappy cuvid can't do 10 bit.

[rodrigo081089]

than there is something wrong with your system. i can only guess but this is the wrong thread for this.

crappy cuvid can't do 10 bit.

I reinstalled the drivers again and now i could make work DXVA. When i open a 720p or 1080p it says H/W decoding, but when i open a 4k video it´s only software decoding.

Maybe someone know how to change that?


Thanks and sorry for the offtopic

[wanezhiling]

MPC? MPC-HC or MPC-BE? Install their latest version from here: MPC-HC (https://nightly.mpc-hc.org/), MPC-BE (http://mpc-be.org/builds/)
And then reset all settings (You know how to do that right?:D), ok now 4k h/w decoding is surely back. :)